Blue Twilight

Pythos is an offline, local-first voice assistant built around Gemma 4 models. It helps users complete everyday tasks through natural conversation while keeping the core assistant experience under the user’s control instead of depending on a cloud AI service. The project explores what a practical personal assistant can look like when privacy, responsiveness, and ownership are treated as core design goals rather than afterthoughts. For the hackathon, we built a working desktop prototype with voice interaction, local reasoning, and tool use, showing that useful AI assistance can run close to the user while still feeling capable and natural. A lot of the work was in the engineering details behind the assistant: routing simple requests away from the speed model, handling multi-step commands, recovering from speech transcription errors, keeping tool calls grounded so the assistant does not claim actions it never performed, and building fallbacks for messy platform APIs like macOS Calendar and local app launching. We also added adaptive model behavior, local/offline-friendly architecture, and test coverage around the failure cases we found during live use, which made the project more than a wrapper around a model. Finally, we used Gradium for the text-to-speech and speech-to-text, and the user can optionally provide their OpenRouter API key to use the Gemma 4 31B model.